Hi researchers, this is a question about risk prediction model. As we know, after running the logistic regression model, we get a set of "coefficients" (beta) of the predictors, and we can use them to calculate the probability (risk) of the event. Also, we can transform these coefficients to some point scores (eg, round them into integers) to improve simplicity and facilitate clinical usage. Besides the coefficients (beta), we can also obtain odds ratios (ORs) of the predictors, and I saw a bioinformatician author used ORs to derive risk scores, but in his paper, the author stated that he used coefficients to derive the risk scores. Did the author make a mistake? Thank you!
@ Olena Polyakova, the study is to develop a risk prediction model using logistic regression among patients who were retrospectively collected and whose outcomes(events) retrospectively confirmed. The outcome variable (dependent variable) is binary, and there are about ten predictors (independent variables). After the statistician run the logistic regression, he obtained the coefficients and ORs of the predictors. Then, he rounded the ORs to integers to derive the point risk scores (which is easier to use in clinical settings), but in the manuscript, this statistician wrote that he rounded the coefficients to derive the risk scores.
Exist the method of rounding of regression coefficient and multiplying it with a 10 (Moons K.G., Harrell F.E., Steyerberg E.W. Should scoring rules be based on odds ratios or regression coefficients? In: J Clin Epidemiol. 2002, vol.55, nr.10, p.1054-1055)
Hi Ning,
Technically speaking, a logistic regression model consists only of the beta coefficients and the constant term. Everything else is derived from that, including the odds ratios which are exponentiated coefficients as you have mentioned. So it is not wrong to say that the simplified scoring rubric that was created by using the coefficients; first into odds ratios, then with simple arithmetic manipulation (i.e. rounding off). As you might already know, many predictive tools that are routinely used in clinical practice are derived this way e.g. MASCC score for febrile neutropenia in chemotherapy exposed patients, CHADS for stroke in atrial fibrillation, LACE index for readmission risk prediction.
in this day and age however, it is unnecessary to do this type of rounding off. It can undesirably blunt the model and could reduce its discriminative properties. one could quite easily develop a web tool / calculator to obtain predicted probabilities of the outcome which would preserve the accuracy. But then again, it really depends on the end-user preference and tolerance of false positives/negatives.
As far as I am concerned, how the model is derived is perhaps not so crucial - there are a million ways to develop prediction models. What's more important is whether rigorously validation was performed on independent population. The results should be reported as per TRIPOD recommendations. I've attached the paper here. If implementation is what you are looking for, I've attached another one you might find useful.